Re: [PATCH RFC net-next 0/2] net:sched: Introduce tc flower2 classifier based on PANDA parser in kernel

From: Tom Herbert <hidden>
Date: 2021-09-23 04:35:12

On Wed, Sep 22, 2021 at 8:25 PM John Fastabend [off-list ref] wrote:

Tom Herbert wrote:

quoted

On Wed, Sep 22, 2021, 6:29 PM John Fastabend [off-list ref]
wrote:

quoted

Tom Herbert wrote:

quoted

On Wed, Sep 22, 2021 at 2:41 PM John Fastabend [off-list ref]

wrote:

quoted

Tom Herbert wrote:

quoted

On Wed, Sep 22, 2021 at 11:00 AM Simon Horman <

simon.horman@corigine.com> wrote:

quoted

On Wed, Sep 22, 2021 at 10:28:41AM -0700, Tom Herbert wrote:

quoted

On Wed, Sep 22, 2021 at 8:49 AM Simon Horman <

simon.horman@corigine.com> wrote:

quoted

On Wed, Sep 22, 2021 at 07:42:58AM -0700, Tom Herbert wrote:

quoted

On Tue, Sep 21, 2021 at 9:46 PM Jiri Pirko [off-list ref]

wrote:

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
Wed, Sep 22, 2021 at 06:38:20AM CEST,

xiyou.wangcong@gmail.com wrote:

quoted

On Thu, Sep 16, 2021 at 1:02 PM Felipe Magno de Almeida
[off-list ref] wrote:

quoted

The PANDA parser, introduced in [1], addresses most of

these problems

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
and introduces a developer friendly highly maintainable

approach to

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
adding extensions to the parser. This RFC patch takes a

known consumer

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
of flow dissector - tc flower - and  shows how it could

make use of

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
the PANDA Parser by mostly cutnpaste of the flower

code. The new

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
classifier is called "flower2". The control semantics

of flower are

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
maintained but the flow dissector parser is replaced

with a PANDA

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
Parser. The iproute2 patch is sent separately - but

you'll notice

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
other than replacing the user space tc commands with

"flower2"  the

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
syntax is exactly the same. To illustrate the

flexibility of PANDA we

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
show a simple use case of the issues described in [2]

when flower

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
consumes PANDA. The PANDA Parser is part of the PANDA

programming

quoted

model for network datapaths, this is described in
https://github.com/panda-net/panda.

My only concern is that is there any way to reuse flower

code instead

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
of duplicating most of them? Especially when you

specifically mentioned

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
flower2 has the same user-space syntax as flower, this

makes code

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
reusing more reasonable.
Exactly. I believe it is wrong to introduce new classifier

which would

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
basically behave exacly the same as flower, only has

different parser

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
implementation under the hood.

Could you please explore the possibility to replace

flow_dissector by

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
your dissector optionally at first (kernel config for

example)? And I'm

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
not talking only about flower, but about the rest of the

flow_dissector

quoted

quoted
quoted
quoted
quoted
quoted
quoted
quoted
users too.
+1
Does the existing BPF flow dissector not work for some reason? If its

purely

quoted

a hardware mapping problem, couple questions below.

Hi John,

eBPF in its current form is un-acceleratable is a primary problem,

We have no disagreement here. Offloading a general purpose instruction
set into an architecture (switch, etc.) that looks nothing like this
is a losing game.

quoted

however an eBPF flow dissector would still have the same issues in
complexity and manageability that the kernel flow dissector. PANDA

PANDA is a DSL and runtime, similarly P4 is another DSL and runtime
environment. P4 can compile to BPF, PANDA can as well. I don't think
PANDA or P4 belong in kernel. We can argue about preferred DSLs but
I don't think that is a @netdev @bpf concern.

quoted

pretty much can address that since the same source code for flow
dissector used in tc-flower could be compiled into eBPF and used with
TC (that gives extensibility at some incremental performance
degradation). There is an interesting caveat there in that to satisfy
verifier we had to break up the parser to avoid being flagged for
complexity. The parser code however already started with a restricted
structure that enforces rules that prevent the issues the verified is
checking for, when we get to verifier though it doesn't have any that
context and hence views it as general purpose code. Introducing domain
specific constructs into eBPF, like the byte code I mentioned, would
simplify the verifier in that case.

We have some type specific constructs already. For example we know
certain helpers will return a max value, etc. and can avoid extra
bounds checks. If we can safely make more assumptions based on
the types of the program and/or types of variables lets do it. I'm
all in for optimizations in the verifier side.

If you have specific complexity problems we can work to solve those
either in the clang backend or in verifier itself. Solving these
will help all use cases so again happy to help here.

quoted

quoted
quoted
quoted
quoted
quoted
quoted
Hi Jiri,

Yes, the intent is to replace flow dissector with a parser

that is

quoted

quoted
quoted
quoted
quoted
quoted
quoted
more extensible, more manageable and can be accelerated in

hardware

quoted

quoted
quoted
quoted
quoted
quoted
quoted
(good luck trying to HW accelerate flow dissector as is ;-)

). I did a

quoted

presentation on this topic at the last Netdev conf:
https://www.youtube.com/watch?v=zVnmVDSEoXc. FIrst

introducing this

quoted

with a kernel config is a good idea.

Can we drop hyperbole? There are several examples of hardware

that

quoted

quoted
quoted
quoted
quoted
quoted
offload (a subset of) flower. That the current kernel

implementation has

quoted

quoted
quoted
quoted
quoted
quoted
the properties you describe is pretty much irrelevant for

current hw

quoted

offload use-cases.

Simon,

"current hw offload use-cases" is the problem; these models

offer no

quoted

extensibility. For instance, if a new protocol appears or a user

wants

quoted

to support their own custom protocol in things like tc-flower

there is

quoted

no feasible way to do this. Unfortunately, as of today it seems,

we

quoted

are still bound by the marketing department at hardware vendors

that

quoted

pick and choose the protocols that they think their customers

want and

quoted

are willing to invest in-- we need to get past this once and for

all!

quoted

IMO, what we need is a common way to extend the kernel, tc, and

other

quoted

applications for new protocols and features, but also be able to

apply

quoted

that method to extend to the hardware which is _offloading_

kernel

quoted

quoted
quoted
quoted
quoted
functionality which in this case is flow dissector. The

technology is

quoted

there to do this as programmable NICs for instance are the rage,

but

quoted

we do need to create common APIs to be able to do that. Note this
isn't just tc, but a whole space of features; for instance, XDP

hints

quoted

is nice idea for the NIC to provide information about protocols

in a

quoted

packet, but unless/until there is a way to program the device to

pull

quoted

out arbitrary information that the user cares about like

something

quoted

from their custom protocol, then it's very limited utility...

Vendors have the ability to code up arbitrary hints today. They just
haven't open sourced it or made it widely available. I don't see how

Vendors can do this, but can I do this as a user? If I have a custom
protocol can I write some code for that without calling my vendor and
can I use that with tc-flower offload as easily as any other protocol?
I believe this is what is needed.

I agree it would be great for the vendors to expose this, but I don't
see how flower2 gets us there yet. Said vendors could give us the
tools needed now, but haven't.

quoted

a 'tc' interface would help with this. I suspect most hardware could
prepend hints or put other arbitrary data in the descriptor or

elsewhere.

quoted

The compelling reason to open source it is missing.

I'm not sure about that. It's not so much a question of the mechanisms
to convey the data (although being restricted to just a few bytes in a
receive descriptor for XDP Hints is an obvious limitation), the

We have multi-buffer support coming so we could use pages of data
for metadata if we wanted fairly easily once that lands. We could
even have hardware DMA the metadata into a page and just add that
to the frag list without any copying.

quoted

problem is the generation of the hints themself. For instance, suppose
I want to extract the QUIC connection identifier as an XDP Hint? To do
this we would need the device to be able to parse in UDP payload and
extract the connection identifier. I don't believe that capability is
widespread yet, vendors have parsers but generally they have not made
them user programmable. And even if they did do that, then we need a
common interface to program that functionality-- proprietary APIs are
right out due to risk of vendor lockin. I imagine the P4 advocates
think they are the solution, but then we have to absorb the
maintenance cost of yet another domain specific language and P4 isn't
part of the kernel so it really isn't offloading kernel functionality.
This is the gap the PANDA fills.

I don't think P4 or Panda should be in-kernel. The kernel has a BPF
parser that can do arbitrary protocol parsing today. I don't see
a reason to add another thing on the chance a hardware offload
might come around. Anyways P4/Panda can compile to the BPF parser
or flower if they want and do their DSL magic on top. And sure
we might want to improve the clang backends, the existing flower
classifier, and BPF verifier.

BPF knows about skbs and its attach points. If it can assume
bounds lengths or other simplifying constraints automatically
lets do it.

Without hardware support I can't see any advantage to flower2.


John,

Please look at patch log, there are number of problems that have come up
flow dissector over the years. Most of this is related to inherent
inflexibility, limitations, missing support for fairly basic protocols, and
there's a lot of information loss because of the fixed monolithic data
structures. I've said it many times: skb_flow_dissect is the function we
love to hate. Maybe it's arguable, bit I claim it's 2000 lines of spaghetti
code. I don't think there's anyone to blame for that, this was a
consequence of evolving very useful feature that isn't really amenable to
being written in sequence of imperative instructions (if you recall it used
to be even worse with something like 20 goto's scattered about that defied
any semblance of logical program flow :-) ).

OK, but if thats the goal then shouldn't this series target replacing the
flow_dissector code directly? I don't see any edits to ./net/core.

John,

Well this is just an RFC ;-). We'll follow with the patch series that
more directly illustrates replacing kernel flow dissector.

quoted

The equivalent code in PANDA is far simpler, extensible, and maintainable
and there are opportunities for context aware optimizations that achieve
higher performance (we'll post performance numbers showing that shortly).
It's also portable to different environments both SW and HW.

If so replace flow_dissector then I think and lets debate that.

My first question as a flow dissector replacement would be the BPF
flow dissector was intended to solve the generic parsing problem.
Why would Panda be better?

To be clear PANDA is a programming model for the datapath not an
execution environment-- I view it as complementary to kernel, eBPF,
DPDK, etc. Similar to how we are suggesting to replace the kernel flow
dissector with a PANDA derived flow dissector, we can implement a BPF
flow dissector in PANDA as well. This should be as simple as
recompiling the parser we're using in this patch for eBPF (we'll do
that experiment). As I mentioned the immediate benefits are a
simplified code base and improved program structure in the source
source. With a context aware optimizing compiler we can output C code
that has higher performance than the equivalent handwritten imperative
code. The longer term benefits are that the same exact parser code is
now accleratable and a more user-friendly interface for users adding
their own custom protocols to generic tc-flower.

My assumption here is that BPF should
solve the generic parsing problem, but as we noted isn't very
friendly to HW offload.

Right, not friendly to HW offload but that's also correlated to not
being very friendly programmability in general.

So we jumped immediately into HW offload
space. If the problem is tc_flower is not flexible enough
couldn't we make tc_flower use the BPF dissector?

Yes, in the software model we would use a BPF dissector, I'm just
saying that that for generic tc-flower we would derive that from PANDA
which is going to be better performance and easier to make the changes
in TC flower application for dynamic CLI to use PANDA parser. Without
any additional considerations, all we're really doing is rewriting the
BPF and kernel flow dissector in a simpler model but retaining the
same functionality (although there is a nice advantage that both
instances can be derived from the same code base).

That should
still allow tc flower to do its offload above the sw BPF dissector
to hardware just fine.

Right but we need something to offload, and it really has to be
identical functionality to what kernel does. While we can't offload an
eBPF flow dissector, we can offload a flow dissector which starts from
an expression in PANDA and is also the basis for an eBPF one kernel
uses.

I guess my first level question is why did BPF flow dissector
program not solve the SW generic parsing problem. I read the commit
messages and didn't find the answer.

I do think eBPF is key to providing a generic software parser, doing
an eBPF parser in PANDA has the advantages I mentioned. This does lead
to a more profound question with regards to the kernel flow dissector
we have today in the kernel: can the flow dissector be completely
removed and replaced with just an eBPF hook that is populated with a
default parser program as part of kernel compilation? I believe the
major reason to have a kernel flow dissector instead of a default eBPF
flow dissector is for performance, presumably native kernel code is
faster than eBPF code in kernel. But if the eBPF code is actually
faster, like in PANDA where we can optimize the code with context,
then the performance issue might be moot. So now, instead of eBPF
being used as an optional hook, it's actively replacing some core
kernel functionality from boot (hence we're starting to rewrite the
kernel itself in eBPF which is why I call it profound). That is
potentially a big win for extensibility, flexibility, and
accelerability also. I think this is something we should explore, but
maybe take an incremental approach to get there by first moving the
kernel flow dissector to a simpler model.

Tom

.John

quoted

Tom

quoted

even if we had support the amount of firmware logic going into taking
that parse graph into hardware makes me a bit nervous. I would
prefer to see a proper compiler that generates the microcode and
*CAM table entires necessary to implement said switch logic from
any DSL although I prefer P4. Sure that compiler wont be in
kernel most likely, but I want transparency and I want tools
to configure my software how I want it. Neither requries a
kernel shim.

quoted

Then the flwo is fairly straight forward the XDP program reads the
hints. Then if the rest of the stack needs this in the skb we have
the hash and skb extensions.

Again that's the backend mechanism, the frontend problem is the
generation of the hints data itself.

Sure, but my point is the mechanism exists today to generate
arbitrary hints and no vendor has stepped up and provide
code to do it. The only conclusion I can come up with is its
not valuable on their side to do it.

quoted

... the NIC could run a BPF program if its programmable to that

extent.

quoted

Simon,

True, but that implies that the NIC would just be running code in one
CPU instead of another-- i.e., that is doing offload and not
acceleration. Hardware parses are more likely to be very specialized
and might look something like a parameterized FSM that runs 10x

faster

quoted

than software in a CPU. In order to be able to accelerate, we need to
start with a parser representation that is more declarative than

Agree, but I don't see how configuration of this hardware makes sense
over 'tc'. This is likely to require compiler tools to generate the
microcode or *CAM entries running on the hardware. Having 'tc' run
a few link, add-header commands that can be converted into reconfigured
hardware (thats not just a CPU on a NIC) seems like we will be asking
a lot of firmware. Probably too much for my taste, fixing bugs in
firmware is going to be harder then if vendors just give us the
compiler tools to generate the parsing logic for their hardware.

Show me the hardware that can create new parse trees using flower2,
do they exist?

Working on it :-)

OK I think flower2 should wait for the hardware then.

quoted

If not the flow is like this,

 0. build new parse graph and hardware logic using DSL (P4 or

otherwise)

quoted

 1. apply blob output from 0 onto hardware
 2. build out flower2 graph
 3. flower2 populates hardware but hardware already got it from 0?

I'm missing the point here?

quoted

imperative. This is what PANDA provides, the user writes a parser in

quoted

declarative representation (but still in C). Given the front end
representation is declarative, we can compile that to a type of byte
code that is digestible to instantiate a reasonably programmable
hardware parser. This fits well with eBPF where the byte code is
domain specific instructions to eBPF, so when the eBPF program runs
they can be JIT compiled into CPU instructions for running on the
host, but they can be given to driver that can translate or JIT
compile the byte code into their hardware parser (coud JIT compile to
P4 backend for instance).

I'm not sure hardware exists that can or will take arbitrary 'tc'
commands and build a graph of a new protocols? Also we already have
a SW path for arbitrary flow parser with BPF so I see no reasons
we need to support yet another one. Even PANDA could produce BPF
codes for sofwtare and hardware codes to program the hardware so
why is this needed?

Hardware parsers are not general purpose CPUs, but specialized engines
that expect input in a declarative representation of an annotated

Of course.

quoted

parse graph that includes parameterized functions needed to determine
next protocol and length. If we compile PANDA parser, which is in
declarative representation, into eBPF then we've lost the structure
and essence of the problem so that we can't accelerate the parser in
specialized hardware. The alternative I'm suggesting is to extend eBPF
to include the declarative representation of the parser. So when the
program is downloaded to the kernel it can run in CPU by compiling to
host ISA, the kernel can then give the program to the device that in
turn can instantiate it in the hardware.

This feels like multiple levels of abstraction for no reason.
Why not just open up the device and give it the declaration directly.
I know some people disagree, but why are we creating all this code
when the entire purpose can be done by simply passing the info to
the hardware directly. I see no reason BPF should pick up extra
complexity just so some hardware device can take it and convert
it out of BPF and run some annotations.

quoted

A critical aspect of this method also eliminates any sort of side
channel or kernel bypass to program the device, which means the device
programming is visible to the kernel hence there is no ambiguity as to
what the device is actually doing (this lack of transparency is why
protocol specific offloads like LRO have failed to get traction). This
visibility is necessary if we are ever to go beyond just getting a few
disposable hints from the device and are getting actual operational
data for the stack to consume (like accelerated TXDP where we can jump
straight to TCP receive processing routine from the driver because the
device was able to perform all the stateless TCP and IP processing
before giving the packet the host). Generally, I view this direction
as extending the boundary of the stack into the device.

Agree we need visibility into what the hardware is doing if we want
complex offloads. But, I don't see why we need to complicated the
software stacks to support the offload.

Anyways IMO we need hardware support to make heads or tails of this.

quoted

Tom

quoted

Also yes I saw the netdevconf but still missed the point Sorry.

Thanks
.John

`h`	back out one level
`j`	next message in thread
`k`	previous message in thread
`l`	drill in
`Esc`	close help / fold thread tree
`?`	toggle this help